Call logs, artifacts and outcomes

Open Observe → Call logs for organisation calls, or the assistant's Logs tab for its activity. Analysis summarizes recent calls and measured performance; Reports provides the available reporting views.

Investigate a call

Check its source, state, end reason and assistant revision. Then inspect stage timing, reported usage, knowledge citations and tool activity. Compare the intended release with the version actually used.

Typical investigations include a provider failing at startup, slow transcription, missing public evidence, declined tool approval or an ordinary caller disconnect. Use troubleshooting to work from the symptom to the affected stage.

Older calls can lack measurements. Missing usage or price information is unknown, not proof of zero usage. Human quality reviews are separate from measured latency or recognition error.

Enable artifacts intentionally

In Advanced, configure transcripts, audio recording and structured outputs independently. Browser recording and follow-up analysis need the separate consent choice; phone calls request spoken consent before capture. Declining consent follows the configured call policy.

Artifacts remain scoped to the organisation and authorized management access. Retention removes stored artifacts and outputs. Persistent facts and collected flow fields have separate settings.

If a transcript ends early, analysis may cover only part of the conversation. Recording and outcomes are not guaranteed merely because a call log exists.

Define a useful outcome

Create a named structured output with clear extraction instructions and a bounded JSON Schema. For example, a support result can distinguish completion from missing information:

{
  "type": "object",
  "properties": {
    "requestSubmitted": { "type": "boolean" },
    "needsFollowUp": { "type": "boolean" },
    "summary": { "type": "string", "maxLength": 500 }
  },
  "required": ["requestSubmitted", "needsFollowUp", "summary"],
  "additionalProperties": false
}

Instruct the analysis to mark submission only when the conversation contains confirmed tool success. Outputs are extracted analysis, not independent proof that the external system changed. Check the destination's record for important actions.

Post-call analysis runs asynchronously and can consume model credits. Review failed or partial jobs and webhook deliveries before treating a missing result as a negative outcome.